Insights
Essays and analysis on how decisions get made with AI — shorter and more frequent than our formal publications. For reports, briefs, and working papers, see the research library.
We attach confidence values to claims, so we owe readers a definition: a stated confidence is a promise about frequency — of the claims we mark 0.86, about 86 in 100 should turn out true. Here is the standard we hold those numbers to, and where it comes from.
Enpath is designed as an institution, not a startup — meant to be trusted long after the people who founded it are gone.
The analytical power behind a serious decision has always belonged to those who could afford it. That is the inequality we want to end.
The world we end up in is not something that happens to us. It is the sum of decisions — and decisions can be made more thoughtfully.
When any machine can give a confident answer, a confident answer stops being worth much. Here is what becomes valuable instead.
How to build AI that widens what a person can consider, then steps back — instead of quietly making the decision for them.
A short, practical discipline for anyone who has to act on machine-generated analysis: five checks, two minutes, before you decide.
The world is rich in frameworks — development goals, climate agreements, governance principles — and poor in the translation layer that turns them into a decision a city can actually take.
Governments spend around 13 trillion dollars a year through public contracts. Every one of those contracts is a decision — and the evidence shows that making those decisions inspectable changes what happens to the money.
Forecasts have improved faster than the decisions they are supposed to trigger. The gap between a warning issued and an action taken is a decision problem — and it is where lives are actually saved or lost.
In much of the world, distrust of institutions is the rational default. Systems built for that reality — verifiable by design, not reliant on goodwill — end up more trustworthy everywhere.
Much of the world governs without reliable data — missing censuses, unmeasured informal economies, indicators that simply do not exist. Honest decision tools must carry that silence, not paper over it.
Language models sound equally sure when they are right and when they are wrong. The published research on hallucination and calibration shows why confidence is the wrong signal to trust — and what to ask for instead.
A short definition of the term we use most — and how it differs from analytics, dashboards, and AI that simply answers questions.
A simple, public test for whether a recommendation is trustworthy enough to act on — and why we hold our own work to it first.